ENDEX: Endorsement Index — How language models qualify when citing in generative AI engines

This study investigates how generative AI engines qualify the establishments they cite, a phenomenon operationalized by the ENDEX (Endorsement Index) construct. Based on 2,754 raw responses collected via API (Google Gemini, Anthropic Claude, and OpenAI GPT) about 51 establishments in Campos do Jordao (SP), analyzed by a deterministic Python parser (no LLM), each response was classified into four categories: endorsement, hybrid, neutral, and none. The ENDEX-class distribution varies systematically by engine, by segment, and by query type, and epistemic disclaimer detection reveals directional caution in Claude (65.0% of branded queries contain a disclaimer, against 2.0% of problem queries). The ENDEX x QSA articulation defines an operational matrix with four quadrants, each associated with a distinct GEO strategy. Version 2 (erratum and revision). This version corrects and strengthens v1 after re-auditing the data: Number-by-number audit: every value was re-checked against the source database, with full confirmation (100%). Six reference corrections: the DOI of the QSA reference (Telles, 2026b) pointed, in v1, to a different paper — corrected in both languages; four references were added; and the page numbers of Ouyang et al. (2022), Radlinski and Craswell (2017), and Aggarwal et al. (2024) were corrected via Crossref. Declaration of AI use added, for methodological transparency. The central thesis (language models do not list neutrally — they qualify when they cite, and ENDEX measures that endorsement dimension) remains and is reinforced.